Model Updating with Neural Network Based on Component Model Synthesis

Structural health monitoring usually depends on an accurate finite element (FE) model. Due to the complexity of the structures of long-span bridges, the initial FE model usually needs updating to reduce model errors and improve prediction accuracy. Neural network, relying on its strong ability of pattern matching, has gradually received more attention in the research fields of structural model updating. However, for large-scale complex structures, the amount of degrees of freedom of the model and the parameters need to be updated is huge. When using a single neural network for model updating, the required set of training data will be extensive to ensure the density of training samples, which leads to low efficiency or even infeasibility of network training. In this paper, a method of model updating with neural network method based on Component Model Synthesis (CMS) is proposed. In the proposed method, the sizeable full structure model is first divided into several substructures, and then each substructure is updated by a small neural network, respectively. After that, the updated substructures are assembled by the Craig-Bampton method, where the information of updating parameters from all substructures to the original complete structure, resulting in an updated full structural model. The feasibility and effectiveness of the proposed method are verified by a numerical simulation example of a FE model updating of a plane truss.